Hardware support for conversion between integer and floating-point data
Techniques are disclosed relating to conversion operations in the context of integer and floating-point processor operations. In some embodiments, floating-point pipeline circuitry is configured to execute a single instruction to convert an N-bit integer value to an M-bit floating-point result. To execute the instruction, source modifier circuitry may generate an intermediate M-bit representation based on the N-bit integer value. Fused multiply-add circuitry may perform a fused multiply-add operation to generate the M-bit floating-point result, where the fused multiply-add operation operates on: the intermediate M-bit representation, a quantization scale factor value indicated by the instruction, and a zero-point value indicated by the instruction. In some embodiments that support down-conversion, the floating-point pipeline circuitry is also configured to execute a single instruction to convert an M-bit floating-point value to an N-bit integer result value.
1. An apparatus, comprising:
floating-point pipeline circuitry configured to execute a single instruction to convert an N-bit integer value to an M-bit floating-point result, the floating-point pipeline circuitry comprising:
source modifier circuitry configured to generate an intermediate M-bit representation based on the N-bit integer value; and
fused multiply-add circuitry configured to perform a fused multiply-add operation to generate the M-bit floating-point result, wherein the fused multiply-add operation operates on:
the intermediate M-bit representation;
a quantization scale factor value indicated by the instruction; and
a zero-point value indicated by the instruction.
2. The apparatus of claim 1 , wherein the quantization scale factor and the zero-point value are M-bit floating-point representations.
3. The apparatus of claim 1 , wherein N is eight and M is thirty-two.
4. The apparatus of claim 1 , wherein the N-bit integer value is included in a packed matrix tile being summed prior to a multiplication by matrix multiply acceleration hardware.
5. The apparatus of claim 1 , wherein the floating-point pipeline circuitry is further configured to use the source modifier circuitry and fused multiply-add circuitry to execute a single instruction to convert two N/2-bit integer values to Q-bit floating-point results.
6. The apparatus of claim 1 , wherein:
the floating-point pipeline circuitry further includes destination modifier circuitry; and
the floating-point pipeline circuitry is further configured to execute a single instruction to:
perform a fused multiply-add operation using the fused multiply-add circuitry to generate an M-bit floating-point intermediate value, wherein the fused multiply-add operation operates on:
an input M-bit floating-point value;
a reciprocal of a quantization scale factor, wherein the reciprocal is indicated by the instruction; and
a quotient of a zero-point value and the quantization scale factor, wherein the quotient is indicated by the instruction; and
modify the M-bit floating-point intermediate value using the destination modifier circuitry to generate an N-bit integer value.
7. The apparatus of claim 6 , wherein the destination modifier circuitry is configured to clamp the intermediate value to a representable range of the N-bit integer value.
8. The apparatus of claim 1 , further comprising:
matrix multiply acceleration hardware configured to:
read input registers that include matrix data;
perform matrix multiplication to generate integer matrix entry results;
convert the integer matrix entry results to floating-point representations prior to writing the matrix entry results to destination registers; and
store the floating-point representations in destination registers as matrix multiplication result data.
9. The apparatus of claim 1 , wherein the apparatus is a computing device that further includes:
a central processing unit;
a display; and
network interface circuitry.
10. The apparatus of claim 1 , wherein:
the N-bit integer value is packed with one or more other N-bit integer values in a storage circuit configured to store M bits; and
load circuitry of the apparatus is configured to access both the N-bit integer value and the one or more other N-bit integer values in a single load operation.
11. The apparatus of claim 1 , wherein the apparatus includes:
a plurality of single-instruction multiple-data pipelines configured to execute instructions; and
fixed-function circuitry configured to control the single-instruction multiple-data pipelines to perform operations for at least one of the following types of programs:
graphics shader programs; and
machine learning programs.
12. A method, comprising:
executing, by a computing system, a single instruction to convert an N-bit integer value to an M-bit floating-point value, wherein the executing includes:
generating an intermediate M-bit representation based on the N-bit integer value; and
performing a fused multiply-add operation to generate the M-bit floating-point value, wherein the fused multiply-add operation operates on:
the intermediate M-bit representation;
a quantization scale factor value indicated by the instruction; and
a zero-point value indicated by the instruction.
13. The method of claim 12 , wherein the quantization scale factor and the zero-point value are M-bit representations.
14. The method of claim 12 , wherein the N-bit integer value is included in a packed matrix tile being summed prior to a multiplication by matrix multiply acceleration hardware.
15. The method of claim 12 , further comprising:
executing, by the computing system, a single instruction to:
perform a fused multiply-add operation to generate an M-bit floating-point intermediate value, wherein the fused multiply-add operation operates on:
an input M-bit floating-point value;
a reciprocal of a quantization scale factor, wherein the reciprocal is indicated by the instruction; and
a quotient of a zero-point value and the quantization scale factor, wherein the quotient is indicated by the instruction; and
modify the M-bit floating-point intermediate value to generate an N-bit integer value.
16. The method of claim 15 , wherein the modifying includes clamping the intermediate value to a representable range of the N-bit integer value.
17. The method of claim 12 , further comprising:
reading, by matrix multiply acceleration circuitry, input registers that include matrix data;
performing, by the matrix multiply acceleration circuitry, matrix multiplication to generate integer matrix entry results;
converting the integer matrix entry results to floating-point representations prior to writing the matrix entry results to destination registers; and
storing the floating-point representations in destination registers as matrix multiplication result data.
18. A non-transitory computer-readable medium having instructions of a hardware description programming language stored thereon that, when processed by a computing system, program the computing system to generate a computer simulation model, wherein the model represents a hardware circuit that includes:
floating-point pipeline circuitry configured to execute a single instruction to convert an N-bit integer value to an M-bit floating-point value, the floating-point pipeline circuitry comprising:
source modifier circuitry configured to generate an intermediate M-bit representation based on the N-bit integer value; and
fused multiply-add circuitry configured to perform a fused multiply-add operation to generate the M-bit floating-point value, wherein the fused multiply-add operation operates on:
the intermediate M-bit representation;
a quantization scale factor value indicated by the instruction; and
a zero-point value indicated by the instruction.
19. The non-transitory computer-readable medium of claim 18 , wherein:
the floating-point pipeline circuitry further includes destination modifier circuitry; and
the floating-point pipeline circuitry is further configured to execute a single instruction to:
perform a fused multiply-add operation using the fused multiply-add circuitry to generate an M-bit floating-point intermediate value, wherein the fused multiply-add operation operates on:
an input M-bit floating-point value;
a reciprocal of a quantization scale factor, wherein the reciprocal is indicated by the instruction; and
a quotient of a zero-point value and the quantization scale factor, wherein the quotient is indicated by the instruction; and
modify the M-bit floating-point intermediate value using the destination modifier circuitry to generate an N-bit integer value.
20. The non-transitory computer-readable medium of claim 18 , the circuit further comprising:
matrix multiply acceleration hardware configured to:
read input registers that include matrix data;
perform matrix multiplication to generate integer matrix entry results;
convert the integer matrix entry results to floating-point representations prior to writing the matrix entry results to destination registers; and
store the floating-point representations in destination registers as matrix multiplication result data.